Work

A soil moisture inversion method for potato fields based on the HA-ResUNet++ segmentation model

Year: 2026

Type: article

Abstract: China is the world's largest producer of potatoes, and its growth is highly sensitive to water. Insufficient water during key growth stages can significantly affect yield and quality. Currently, there are issues in drone thermal infrared remote sensing monitoring, such as inaccurate extraction of canopy temperature and low accuracy in soil moisture inversion. This study collected thermal infrared images and ground-based measured data at the experimental base in Zhangbei County, Hebei Province, during the seedling stage, tuber initiation stage, and tuber bulking stage of potatoes. This study proposes a HA-ResUNet++(Hierarchical Attention ResNet34-UNet++) canopy segmentation model, which integrates Deformable Convolution (DCNv4), Convolutional Channel Attention (CBAM), and Contextual Transformer Attention (CoT Attention), and adopts a four-channel input combining pseudo-color and temperature matrices. Experimental results showed that the proposed model improved IoU, mIoU, PA, and F1-Score by 9.58%, 7.31%, 3.98%, and 6.09%, respectively, compared with UNet++, and outperformed several classical segmentation methods. Based on the segmentation results, four canopy temperature determination methods were used for calculating the Crop Water Stress Index (CWSI), and a quadratic polynomial fitting model was constructed for soil moisture inversion. The optimal method achieved R2 values above 0.5 across three growth stages and two potato varieties. The results demonstrate that high-precision canopy segmentation contributes to improve the stability and applicability of soil moisture inversion models, providing a rapid and non-destructive means for precise irrigation in potato fields.

Source: Smart Agricultural Technology

DOI: 10.1016/j.atech.2026.101993